IN-CAR SPEECH RECOGNITION USING DISTRIBUTED MICROPHONES - DSP for In-Vehicle and Mobile Systems - page 302

5.2 Baseline Performance of MRLS
For the evaluation of the baseline performance of MRLS, five recogni-
tion experiments are performed:
CLS - TALK recognition of close-talking speech using close-talking HMM
MRLS SPKER recognition of MRLS output optimized for each speaker
using MRLS HMM
MRLS DR recognition of MRLS output optimized for each driving con-
dition using the MRLS HMM
MRLS ALL recognition of MRLS output optimized for all training data
using MRLS HMM and
DIST recognition of nearest distant microphone speech by the distant
microphone HMM.
The resulting recognition accuracies are listed in Table 19-4, and the
average accuracies over fifteen driving conditions are shown in Figure 19-4.
It is found that MRLS outperforms the nearest distant microphone result
even in “MRLS ALL”, where a set of universal weights are used for all
conditions. This result confirms the robustness of the MRLS to the change
of the location of the noise sources, because the primary noise locations
are different depending on driving conditions. It is also found that the
improvement is greater when the performance of the distant microphone
is lower.
5.3 MRLS Performance with Weight Adaptation
To evaluate the MRLS performance with weight adaptation, optimal
regression weights for the four noise clusters of Section 3 are trained.
Using a 200 ms non-speech segment preceding the utterance, the nearest
prototype of the noise cluster is searched; then the utterance is recognized
after MRLS with the regression weights optimized for the corresponding
noise cluster using the same MRLS HMM. The results of the experi-
ments are shown in Figure 19-5, where the performance of the MRLS
using adaptive regression weights is as high as the results of using the
optimally trained weights for each driving condition. Furthermore, the
MRLS outperforms the MLLR adaptation (five-word supervised adapta-
tion) applied to the close-talking speech [9]. Therefore, the effectiveness
of the proposed method is confirmed.